相变存储器材料设计与多尺度模拟的研究进展

沈雪阳, 褚瑞轩, 蒋宜辉, 张伟

Progress on Materials Design and Multiscale Simulations for Phase-Change Memory

SHEN Xueyang, CHU Ruixuan, JIANG Yihui, ZHANG Wei

图7 根据文献[37,97,118,119]得到的机器学习原子间势的拟合过程及方法

Fig.7 Process and method for fitting machine learning interatomic potential
(a) construction process of machine learning interatomic potential, according to Ref.[97] (

ξ

—process of getting the smooth overlap of atomic positions descriptor, σ_at—smooth parameter, r_cut—specified cutoff, ρ_i (r_cut)—obtained atomic density, q—descriptors, E(q)—energies; the red, blue and yellow spheres represent Ge, Te and Sb atoms, respectively)
(b) sketch of the neural, networks according to Ref.[118] ( G₁, G₂—input vectors describing the atomic configurations,

γ_{j}^{i}

—weight sum of the node values, α^ij —connecting weight parameter, ɛ—potential energy)
(c) schematic of the graph convolutional neural networks^[119]
(d) schematic of the kernel based Gaussian approximation potential, according to Ref.[37] (E_i —potential energy, α_i —regression coefficient, k—SOAP similarity)